Why is AI so Dumb?

Why automating knowledge acquisition offers a path forward.

Topic map of collective reasoning (image by author)

“Why is AI so Dumb?” is the title of the cover story for IEEE Spectrum’s October Issue. In a series of articles, a comprehensive analysis of AI suggests that we are at a significant inflection point for the path forward in AI applications and AI research. The purpose of this article is to explore the conclusions this question poses and suggest a way forward. Are we headed into a new AI winter, or is there a path forward that promises continued growth and value?

I lived through the first AI winter.

As a leader of the first AI IPO (INAI — IntelliCorp) and as co-Chairman of AAAI 1986, I experienced the euphoria of an AI wave that had immense successes. Those stories, lost in the aftermath of a long cold winter, contain hints of a path forward. Nearly every CEO of leading corporations paid attention to and placed an R&D bet on building expert systems and other AI applications. Some were quite successful prototypes that led to broad new application categories such as demand forecasting/pricing, configuration management, diagnostic methods, and fraud detection, to name a few. The promise was substantial, but the costs were higher than the benefit. As a result, the successful application categories evolved into traditional software applications. There was value in the productivity AI technologies offered, but there were roadblocks to broad adoption.

One of the biggest roadblocks to scalability was the fact that knowledge acquisition was a labor-intensive process. However, the difficulty did not stop everyone. I was in Doug Lenat’s office in the mid-80s. His stated mission: build a massive, comprehensive knowledge base. Nearly 40 years later, the project goes on and is finding utility. That said, handcrafting knowledge bases limits scalability.

In the cover story overview, two quotes clarify the current state of affairs:

“Neither symbolic AI projects such as Cyc from Douglas Lenat nor the deep learning advances pioneered by Geoffrey Hinton, Yann Lecun, and Yoshua Bengio have yet produced human-level intelligence.” ¹

“In terms of how much progress we’ve made in this work over the last two decades: I don’t think we’re anywhere close today to the level of intelligence of a 2-year-old child. But maybe we have algorithms that are equivalent to lower animals for perception.” ²

The first comment points to the limitations posed by the current state of affairs: handcrafting AI systems do not scale. The second states that methods based on learning mathematical patterns from data (though genuinely robust and powerful) are inadequate to capture the essence of human intelligence. Finally, Judea Pearl takes the point a step further:

“You are smarter than your data. Data do not understand causes and effects; humans do… The surest kind of knowledge is what you construct yourself.” ³

Pearl’s comment gives a hint to a path forward which we pick up in later paragraphs.

While the awareness of the limitations of AI is increasingly coming to the forefront, researchers involved in the field recognize the limitations. A few years ago, I attended a conference at MIT that marked the launch of MIT Intelligence Quest, a broad initiative to advance human and machine intelligence research. Stanford and Berkeley launched similar programs to deepen cross-disciplinary research in understanding the nature of intelligence (human and artificial). In addition, various research initiatives from NSF and DARPA seek to build better foundations for integrating human cognition with artificial models of intelligence. Recognizing the limitations and the need, what is the path forward for practical applications of AI?

Generalized AI is settling into a long path of fundamental research as signaled by MIT-IQ. What is the best path forward? Can we employ developments in human empowered AI and collective intelligence? I believe the answer is a definitive yes. Still, it requires diving deeply into reframing how we approach AI applications by finding common ground shared by human collaborative learning and deep learning. The essence of a novel approach began in the late 90s.

As the AI winter set in, I took a 5-year recess from AI and became CEO of a company spun out of Apple in 1992. Connect became a platform for eCommerce that went public in the late 90s. Near the end of my time with Connect, a friend came to me with a process he originally called “The Thinking Game.” ⁴ It was a manual process of collaborative learning and convergence initially implemented with index cards. The process and resulting model were inspired by studying the process of scientific discovery. ⁵ By the end of the 90s, we created a web-based collaborative learning platform we trademarked as “the adaptive conversation.” Procter and Gamble and NBC were early adopters of the technology. We were able to show that one could make predictions by learning why a consumer liked a product or a viewer liked (or disliked) a television show.

Work on the adaptive conversation coincided with the publication of The Wisdom of Crowds and The Tipping Point. The first book pointed to what became the science of collective intelligence, and the second covers how emergent patterns grow into more significant market phenomena. Collective intelligence is about how we are smarter together. As discussed in The Tipping Point, complex adaptive systems model early trends that lead to massive impact.

Combining collective intelligence and complex adaptive systems: a new framework for AI

The core technology for a new framework for AI is to acquire knowledge from collections of humans and store that knowledge in a live computational model using a process we call collective reasoning. Specifically, the collective reasoning process builds probabilistic graphical networks that contain the collective intelligence of the participants. Thus, the AI is a kind of super-intelligent meeting facilitator that guides the group through a process that balances discovery with optimizing learning alignment, producing, in the end, a probability and an explanation. Driven by a template that embeds a heuristic model (e.g., a decision rubric or scorecard), the system returns a probabilistic estimate of the outcome with a full explanation of the collective reasoning.

From twelve years of working with leading consumer companies (P&G, LEGO, NBC, Intuit), we learned that learning preferences from bottom-up consumer conversations (e.g., what would you like to experience?) was dependably predictive. The core adaptive learning algorithm could produce a relevant ranking of statements. The ranking accuracy passed the scrutiny of the world’s leading market intelligence professionals both in CPG and media. We also learned that it captured the attention of top management leading to the technology driving the results for two of the five case studies in Reichheld’s NPS book “The One Question.” It created transformative cases in innovation, LEGO being a prime example. The then CEO of LEGO Direct used the technology with the LEGO community to co-innovate what became the Star Wars Imperial Star Destroyer by starting with a simple question: “What do you want to experience when you pour your LEGO’s on the floor?”

Adaptive learning worked. It turned qualitative conversations into quantitatively predictive results, but we had only scratched the surface. We did not understand the underlying reasons for the performance. We did not have any way to process the results other than through services (e.g., no reliable NLP technology), and the result was a report. We needed robust NLP technology to unlock the power of the new framework for integrating human and artificial intelligence.

Integrating human and artificial intelligence

The framework for integrating human and artificial intelligence is described in more detail elsewhere ⁶, and it consists of three interoperable components: 1. A component that learns a probability of relevance for all reasons of the system based on learning alignment of peers using hidden Markov modeling techniques, 2. NLP component that creates a geometric model of the reasoning space, and 3. A probabilistic graphical model of the collective reasoning process. ⁷

The technology development for the above framework commenced in 2014. The objective was to determine if one could take a data-poor asset such as an early-stage startup and predict as an outcome the likelihood of the startup gaining follow-on investing. Survivability is the most significant correlate to a successful return on investment. The frame-based model for predicting fundraising potential was derived from research sponsored by the Angel Capital Association. ⁸ The “data” for this prediction came from a cognitively diverse group of collaborators. Some members had investment experience, some with industry experience, some with entrepreneurial experience, and some with the end-user experience. The collaborating teams were selected based on the domain expertise required for each startup and were anywhere from 12 to 30 individuals. The data sets analyzed over the four years were approximately 150 companies. The result was quite positive >80% that scored above a threshold of 73% went on to gain market attention in the venture investing market. The top-scoring company is currently valued at >20x its value and won two CES awards. All models produced through this process are stored as Bayesian Belief Networks. The knowledge models are persistent computational models of the collective reasoning process. Once created, the models function as mini expert systems. For example, the models for the investment use cases are combinable and helpful in learning how to become better at early-stage investing. ⁹

From those years of experience, we learned that we had a more general knowledge acquisition engine applicable to various areas of market intelligence or knowledge discovery processes associated with decision-making. For example, learning evidence-based alignment in a single-blind collective reasoning process and subject to peer review is essentially a process of knowledge discovery that is highly aligned with how we build models derived from scientific research.

For that reason, a suggested path forward for the next generation of AI technologies is to use collaborative reasoning to reach across disciplines in a form similar to the cross-disciplinary efforts at MIT IQ. Collective reasoning applies to evaluating alternative models and technologies. At a commercial level, collective reasoning applies to assessing the impact potential of technologies in innovation programs and investments that fuel corporate transformation initiatives.

The mathematical and scientific underpinnings of collective reasoning technology are very similar to that of deep learning. Deep learning works as well as it does because it has fundamental roots in cooperative phenomena in physics. Specifically, at the core of deep learning is a model derived from the physics of magnetic behavior. For example, microscopic alignment of electron spin scales to macroscopic magnetic properties via a mathematical formulation called the renormalization group. Deep learning scales how data aligns in patterns from microscopic interactions (e.g., pixel to pixel) to macroscopic properties (“it’s a dog”) by implementing a renormalization group process.

The collective reasoning system we have developed uses similar mathematical formulations but applies it to how humans align their knowledge to formulate a decision or a prediction. By instrumenting the collaboration process, we learn alignment. Specifically, the alignment learning algorithm intelligently samples a shortlist of reasons submitted by other participants seeking to discover areas of alignment. A frame-based knowledge acquisition template guides the system.¹¹

Think of a frame as simply a way to focus on a specific aspect of a decision or prediction. For example, frames for an early-stage investment might be as simple as “What do you think of the market opportunity?” and “What do you think of the team?”. The collective reasoning system will learn the collective reasoning perspective on the market opportunity and the team.

The collective reasoning algorithm balances discovery (allowing participants to think openly and expand their reasoning based on seeing others’ reasons) and optimization (learning alignment).⁷ Similarly, a human facilitator of a group meeting would move from discovery to convergence (get all ideas out and then prioritize). The learning algorithm has methods for learning when to shift attention from discovery to prioritization.

Automating collective reasoning has enormous potential. The system learns rank-ordered preference of reasons in support of a decision or prediction for a group of any size. Thus, collective reasoning has broad applicability from organizational decision-making to predictive market intelligence models.

The first wave of AI focused on handcrafting systems that modeled human intelligence. We put the human mind on the pedestal, examined it, and built models of knowledge and reasoning using logic and heuristics that attempted to reflect and model human expertise. The second wave of AI placed statistical correlation on the pedestal and had great successes in automated pattern recognition and categorization.

The third wave of AI places combined intelligence on a pedestal. The MIT definition of Collective Intelligence says it clearly: “how people and computers can be connected so that — collectively — they act more intelligently than any person, group, or computer has done before. Collective reasoning provides a path forward that promises to lead to that goal. Collective reasoning takes the mathematically sound method of learning the implications of scaling alignment in deep learning to scalably learn alignment in humans, creating computational models of collective human knowledge.

Learning alignment of shared reasoning about a decision or outcome creates a framework for knowledge acquisition that closely mirrors the process of scientific discovery. Furthermore, it is grounded in an underlying principle from physics, cooperative phenomena, that promises a firm foundation for future research and development.

If you would like to be part of a new initiative exploring collective reasoning, please email me at tom@crowdsmart.ai. For commercial use, the technology is licensed as a cloud-based service from CrowdSmart.ai. If your interests are research-related, please make that clear in your email.

1. Eliza Strickland “The Turbulent Past and Uncertain Future of AI” IEEE Spectrum October 2021
2. Yoshua Bengio, founder and scientific director of Mila-Quebec AI Institute
3. Judea Pearl interview The Book of Why Basic Books 2018
4. Brad Ferguson
5. Thomas Kuhn “The Structure of Scientific Revolutions” University of Chicago Press 1970
6. Thomas Kehler Transforming Collaborative Decision Making with Collective Reasoning
7. Patents pending
8. Wiltbank, Robert, and Boeker, Warren, Returns to Angel Investors in Groups (November 1, 2007). Available at SSRN: https://ssrn.com/abstract=1028592 or http://dx.doi.org/10.2139/ssrn.1028592
9. Thomas Kehler Applying Collective Reasoning to Investing
10. Henry Lin, Max Tegmark, David Rolnick “Why does deep and cheap learning work so well?” arXiv:1608.08225
11. Richard Fikes and Tom Kehler “The Role of Frame-based representation in reasoning” CACM Vol 28 No 9